Disney built its own premium format because it could not get a screen.

Before anything else: this study’s own brief had the date wrong

The pre-registered brief said December 19. All six model answers said December 18, 2026, and they were right. December 18 is a Friday; December 19 is a Saturday, and US wide releases open on Fridays.

The error came from a trade article read once and not checked against a second source — a failure of the two-source rule this estate applies to everything except, on that day, its own premise. It is corrected here first, because every other figure on this page is anchored to that date. It is also why the sourcing table further down exists at all.

The registered question, written before any retrieval

When two wide releases collide on a single date, is premium-format screen allocation the scarce resource that decides the outcome — and does demand measured before the date predict which title leads?

The seat this was written for is whoever allocates premium-format screens, and whoever on the other side decides whether to pay into an incumbent format or build one. That second decision is live, recurring and expensive. The December 18 allocation itself was not live when we measured — it had already resolved, and we registered it as open. That is half a hit, and the half that missed is the one that matters.

Every demand instrument says Avengers

Four of the five clear 70%, which was the registered threshold for dominance. The row worth looking at is ticket searches. It is the closest thing here to purchase intent and it is the narrowest gap of the four search measures — 73.2% against 95.3% on the head term. Dune closes the gap as the question moves from awareness to buying. Keyword volumes are advertising bands rather than counts, so that is a direction and not a ratio.

Every answer engine that recommends says Dune

Asked which is bigger, every instrument says Avengers. Asked which to see, the engine that answers at all says Dune, three times out of three. That is not a contradiction. They are two different questions, and the brief registered a measurement that could not tell them apart — which is logged as a defect in the table above rather than smoothed over here.

The commercially useful version of this is not who wins December 18. It is what a 96-to-1 search advantage is worth when the other title holds the only screens that carry a premium ticket price. That has a dated answer and nobody has published a measurement of it.

How the allocation actually happened

Every entry carries its source and how strong that source is. The first one is the one to watch, because of where it ends up.

The response is the finding. A studio that could not get one chain of screens for one weekend launched a certification program for a competing class of screens seven weeks later, and ten weeks after that reported more than 7,500 applications. That is a revealed price for premium-format access, and it is larger than the forgone ticket revenue of a single weekend.

The number both engines gave is in none of the sources they cited

Perplexity recommended Dune three times out of three, and gave a reason each time: the exclusivity runs at least its first three weeks. It cited a real article for it. That article does not contain a duration.

Two durations are in circulation. The shorter one — the opening weekend — traces to IMAX’s own investor presentation. The longer one — three weeks — traces to a post on X, reported by a film blog that hedged it with “apparently”. The engines gave the longer one, and attached it to a citation that supports only the fact of precedence.

This matters beyond one film. A confident number with a real source attached, where the source does not contain the number, is the hardest kind of error to catch — the citation survives a spot check and the figure does not. Anyone taking a competitor’s exclusivity window from an answer engine and putting it in a plan would have taken the unsourced version, twice as long, with a footnote that looked right.

Sixth consecutive replication, and ChatGPT returned zero again

ChatGPT’s open-question citation set shares not one domain across three calls placed in the same minute. Calls 1 and 3 returned one set; call 2 returned an entirely different one. Perplexity has now returned a fully identical citation set on six consecutive measurements, across four subjects and two question shapes. This is a measurement of these questions on these days and not a league table, and we do not publish it as one — but six for six is past the point of hedging into meaninglessness.

Who answers for a film, and it is not the trades

Across twelve calls, neither engine cited Variety, Deadline or The Hollywood Reporter once. On the question of the biggest American film release of December, ChatGPT’s first-position citation on two of three open calls was a Peruvian news site.

One owned domain did win: movies.disney.com was cited at position 1 on two of three head-to-head calls. Compare the games study, where the publisher was cited zero times out of three about its own release date. The difference is the question — this one named both titles. It is also a subdomain, which a measurement matching only the root domain would have recorded as a miss.

The prediction, registered now, resolved December 21

A study about a dated event is worth little if it does not say what it expects before the date. These are written now and will be scored here, in place, whichever way it goes.

What would make this study wrong: if Dune takes the larger opening, then premium-format allocation outweighed a 95-to-5 search advantage, and the demand instruments on this page do not predict opening gross. That is a result worth more than being right, and it will be published as plainly as the other one.

The basis is structural, not measured, and that matters to how much the number is worth. cm_123 separates a prior set from an earlier measurement with the same instrument from one set from a plausible belief. Across 57 priors carrying an explicit basis label, the first stands at 14 for 19 and the second at 18 for 33 — a nineteen-point edge, not an unbroken record. An earlier version of this page said 17 for 17 against 0 for 6; that figure counted an unresolved prediction as a hit and described only the minority of projects that labeled their basis. P4 is set the second way. It is published at 75% anyway, because a prediction withheld until it is safe is not a prediction.

What would make P4 unresolvable rather than wrong: per-screen averages for premium formats are not reliably published. If neither IMAX nor Disney reports screen counts and format grosses for the opening weekend by January 9, 2027, P4 is recorded unresolvable and scored as neither a hit nor a miss. Silence is not a win.

What this study cannot tell you

• Nothing here observes a ticket sold. Search volume is interest, comments are attention, and a model recommendation is a model recommendation. No row in this study measures a booking, a presale or a screen allocation.
• The allocation itself is reported, not inspected. IMAX’s investor presentation is cited by two outlets that read it; the deck was not retrieved and read here. Neither studio confirmed anything to us.
• Keyword volumes are Google’s advertising bands, not counts. Read the ordering. The 95.3% and 73.2% figures are band ratios.
• The trailer comparison is one official trailer per title, four months apart — Dune’s teaser in March, Avengers’ trailer in July. Age is not controlled and comments accumulate, so 86.1% is a floor rather than a point.
• Both engines were read at the platform-default budget tier, on one day, from one location. A different tier or a different day may answer differently.
• No negative control. The brief registered one and none was appointed, because no true negative was available. That is a gap, stated rather than papered over.
• The 75 domestic and 300 international figures are disputed between sources. One reads them as Infinity Vision certified theaters, another as the existing premium-format screens the program draws from. We use the second, cautious reading, and neither is used in a calculation here.

How this page changes

A standing study, not an article. Corrections are published in place and dated, never applied quietly.

Related reading, free and in full

• All standing studies — the index and the method each one follows
• GTA VI: the same citation measurement in a different industry, where the publisher was cited zero times out of three about its own release date
• Your category’s advertising benchmark is noise — 1,633 filers, every rejected row published
• The IP Association Map: when a question names the brand, the brand’s own domain is cited 82% of the time; when it does not, 42%